Data modelling:

Predicting bed availability in intensive care

 

 

An Intensive Care Unit (ICU):

In [5]:
x = Image(filename='icu.JPG', width=600)
display(x)

This is where you go when you are very sick:

  • After major surgery
  • Multiple organ failure
  • Serious co-morbities, infections etc.

Jeremy needs a serious operation.

In [4]:
Image(filename='jeremy.jpg', width=600)
Out[4]:

But on the morning of his operation, the ICU is full.

In [6]:
Image(filename='bed_plan.jpg', width=600)
Out[6]:

So..

In [7]:
Image(filename='where_jez.jpg', width=600)
Out[7]:

Going to go after his operation?

Nowhere.

- His operation is cancelled.

- The hospital receives a penalty fine.

This situation is not uncommon..

In [9]:
Image('news.png')
Out[9]:

What can be done?

1. More intensive care beds (and staff).

2. Better use of the available resources.

Making better use of resources..

In [6]:
Image(filename='bed_plan.jpg', width=600)
Out[6]:

If we could predict when beds will become free:

  • Schedule electives better
  • Improve 'Flow management'

The challenge:

Predict when patients will leave intensive care.

Specifically:

Based on the data collected during the first 48 hours in intensive care, can you predict if a patient will still be undergoing treatment by the end of day X?

The data:

Real world data for 4000 intensive care patients:

  • time series recordings of physiological parameters
  • demographic information
  • outcomes

Physiological parameters:

In [3]:
Image('variables.png')
Out[3]:

For example the Temperature time series:

In [21]:
variable='Temp'
unit='(deg C)'
plot_var(example, variable, unit)

..the Heart rate time series:

In [22]:
variable='HR'
unit='(bpm)'
plot_var(example, variable, unit)

..or the Alkaline phosphatase time series:

In [23]:
variable='ALP'
unit='(IU/L)'
plot_var(example, variable, unit)

..and so on.

And importanlty, the OUTCOME information:

  • Length of stay (in days, always > 2)
  • In-hospital death (0: survivor, 1: died in-hospital)

  • Survival (days)

  • additionally: SAPS-I and SOFA score.

Missing data given by -1

Advice:

  • You should build at least one predictive model or classifier.

  • You should split the data into a training set and a test set. Beware of overfitting, you want performance to generalise.

  • Test the performance of your model using concepts such as Recall, Precision, and if possible AUC (the area under the ROC curve).

  • I would recommend starting with X = 5 days.

Extensions (optional):

  • Are there systematic difference between the training and test sets? What issues might this cause, and what could you do to overcome them?

  • Think about the different reasons for leaving ICU. Can you account for these in your modelling?

  • Try developing a different model to predict the scalar length of stay (LOS). Use this new model to predict which patients will have left ICU by day X. How does this compare to your original results?

Data:

Are available from: https://physionet.org/challenge/2012/ (use 'set A')

Contacts:

  1. Chris McWilliams: chris.mcwilliams@bristol.ac.uk
  2. Alan Champneys: a.r.champneys@bristol.ac.uk